Re: threaded programs report bogus process size ?

[christos%astron.com@localhost (Christos Zoulas)]

In article <20090930124029.GA275595%medusa.sis.pasteur.fr@localhost>,
Nicolas Joly  <njoly%pasteur.fr@localhost> wrote:
>On Wed, Sep 30, 2009 at 01:59:24PM +0200, Nicolas Joly wrote:
>> On Wed, Sep 30, 2009 at 12:16:30AM +0000, Christos Zoulas wrote:
>> > In article <h9u2f5$qap$1%serpens.de@localhost>,
>> > Michael van Elst <mlelstv%serpens.de@localhost> wrote:
>> > >njoly%pasteur.fr@localhost (Nicolas Joly) writes:
>> > >
>> > >>I just noticed that threaded processes may report increasing process
>> > >>size of their life ...
>> > >
>> > >libpthread never frees the pthread data structure and associated
>> > >thread stack when a thread ends but puts it onto a private queue
>> > >("deadqueue") and reclaims it again when a new thread is created.
>> > >
>> > >The reclaim can fail when a new thread is created shortly after
>> > >an old thread ended because of a race with the kernel. As a result
>> > >a new thread is allocated.
>> > >
>> > >So _some_ leakage is allowed by design.
>> > >
>> > >The problem is that the reclaim operation is only tried for the
>> > >thread that ended last (the top of the "deadqueue"). So in a
>> > >tight pthread_create/join loop the leaked threads just pile up.
>> > >
>> > >To correct this the reclaim operation should loop over the queue
>> > >and try to reclaim any of the dead threads.
>> > 
>> > This seems to fix it.
>
>I spoke a little too soon :(
>
>Further tests (+ print loop iteration number) show that at some random
>time, no more threads are created and the program seems stuck in an
>endless loop ...

I'll try to reproduce it.

christos